Where Two Billion Fable 5 Tokens Went
3 min read
Of the 15.6 billion tokens my AI tools moved in the last three months, 2.1 billion went through a single model: Claude Fable 5. That's 13.5% of everything — and because Fable is the most expensive model I run, it's a much larger share of the bill. I wanted to know if that spend was aimed or just spilled. So I went to the transcripts.
Fable is the author, not the shipper
Fable 5 lists at $10 per million input tokens and $50 per million output — double Opus 4.8. You don't hand a model like that your whole workload. You give it the one job where being a notch smarter pays for itself.
For me that job is authoring. In my dev loop, Fable writes the hard first draft — a new worker, a whole skill library, a migration — and a cheaper model (usually Opus 4.8) reviews it, catches the defects, and ships it. Generate with the strong model, verify and land with the workhorse. Fable is the writer's room; Opus is the editor with the red pen and the keys to production.
The token trail confirms the role. Across 12,809 Fable turns, its 10.5 million output tokens are first drafts — code that another pass would tighten, not final answers.
It lives in the shell
Here's the breakdown of Fable's 7,283 tool calls:
| Tool | Calls | Share |
|---|---|---|
| Bash | 4,509 | 62% |
| Read | 1,240 | 17% |
| Edit | 354 | 5% |
| StructuredOutput | 196 | 3% |
| Write | 101 | 1% |
Sixty-two percent Bash — even more shell-heavy than my overall average of 53%. That tracks with the authoring job: writing real software is mostly operating — run the build, read the error, grep the callers, check git status, run the test again. The prose is the small part. The model that authors is the model that spends its life in a terminal.
Where the tokens landed
Fable's output, by project:
- homelab — 5.26M tokens. Half of everything. My infrastructure is where the hardest authoring happens: new workers, the skill libraries, the self-healing plumbing.
- railmind — 743K. A separate project with its own hard first-drafts.
- website — 348K. This site.
- cortex / Coquina — 267K. The memory platform itself.
- The rest scattered across workflow and subagent buckets — Fable fanned out across parallel orchestrated agents, each authoring a slice of a larger job.
No surprises, which is the point. The most expensive model went where the authoring was hardest and nowhere it wasn't.
The cost of being the author
Fable's 13.5% of my tokens cost $4,349 at metered rates — 22% of my total Claude bill for the quarter. Nearly a quarter of the money on an eighth of the tokens. That's not waste; it's the premium priced exactly where I chose to pay it.
The lesson isn't "Fable is expensive, use less." It's that a model this capable is a scalpel, not a faucet. Point it at the first draft of the thing you couldn't have written as fast yourself, then let something cheaper carry it the rest of the way. Two billion tokens went to being the author. Everything downstream — the review, the ship, the overnight grind — ran on models that cost a fraction, because by then the hard part was already on the page.